Obtain real completed vendor responses, including a compliance matrix #25

Closed
opened 2026-08-02 03:24:00 +00:00 by christian · 2 comments
Owner

Question

Nothing to decide here — this is manual work that a decision waits on.

The artifacts research established the RFP side well but could not obtain a single completed
vendor response
. Every candidate returned 403 to automated fetching. Response-side binding is
currently documented only from RFP-side instructions, never from a real artifact.

That gap sits under the most important part of the schema: the map's central insight is that the
response is the artifact worth standardizing, and it is the half designed against imagination
right now.

Routes that need a human, roughly by cost:

  • Public records requests. Awarded proposals are frequently public records — Maine's
    procurement site states submissions "will be considered public records" once awarded. Tarrant
    County and Friendship PCS both have identifiable awarded vendors and are plausible targets.
  • Headless-browser retrieval of the sources that 403'd: Entegra/OMNIA Partners, Cisco
    WSCA-NASPO, the NYC shelf-stable foods bid tab, USDA AMS bid awards.
  • Ask a vendor directly. Anyone in the network who has responded to a large retail or
    institutional RFP and will share a redacted copy.
  • Cooperative purchasing portals, which sometimes publish winning responses as part of the
    contract record.

Done when at least one complete response is in hand — ideally with a filled compliance matrix —
and the requirement-to-response binding has been examined against a real document rather than
inferred.

Record in the resolution: what was obtained, from where, and what it changes about the binding
model.


Parent: #1

## Question Nothing to decide here — this is manual work that a decision waits on. The artifacts research established the RFP side well but **could not obtain a single completed vendor response**. Every candidate returned 403 to automated fetching. Response-side binding is currently documented only from RFP-side instructions, never from a real artifact. That gap sits under the most important part of the schema: the map's central insight is that the **response** is the artifact worth standardizing, and it is the half designed against imagination right now. Routes that need a human, roughly by cost: - **Public records requests.** Awarded proposals are frequently public records — Maine's procurement site states submissions "will be considered public records" once awarded. Tarrant County and Friendship PCS both have identifiable awarded vendors and are plausible targets. - **Headless-browser retrieval** of the sources that 403'd: Entegra/OMNIA Partners, Cisco WSCA-NASPO, the NYC shelf-stable foods bid tab, USDA AMS bid awards. - **Ask a vendor directly.** Anyone in the network who has responded to a large retail or institutional RFP and will share a redacted copy. - **Cooperative purchasing portals**, which sometimes publish winning responses as part of the contract record. **Done when** at least one complete response is in hand — ideally with a filled compliance matrix — and the requirement-to-response binding has been examined against a real document rather than inferred. Record in the resolution: what was obtained, from where, and what it changes about the binding model. --- Parent: #1
christian added the
wayfinder:task
wayfinder:ticket
labels 2026-08-02 03:24:00 +00:00
christian added a new dependency 2026-08-02 03:24:00 +00:00
christian self-assigned this 2026-08-02 03:26:37 +00:00
Author
Owner

Progress — AFK attempt failed, handing to a human

Retried the 403'd sources through a real browser rather than plain HTTP. The bot-challenge theory
was wrong: the documents are not blocked, they are not published.

Searched, nothing found:

  • Sourcewell awarded-supplier proposals (assets.sourcewell-mn.gov returns zero indexed proposals)
  • OMNIA Partners / Region 4 supplier responses
  • Generic hunt for a filled compliance matrix ("compliance matrix" "RFP section" "comply") — every
    hit was the RFP instructing bidders to produce one, never a completed one
  • Louisiana Medicaid MCO RFP 2021, a $9bn award with six named proposers (Aetna Better Health,
    AmeriHealth Caritas, et al.) — LDH publishes the RFP and the proposer list, not the proposals

The finding worth keeping. Solicitations are published because issuers are legally obliged to
advertise them. Responses are commercially sensitive and are published essentially never, even by
issuers whose own RFP says awarded submissions become public records. There is no public response
corpus, in any vertical.

Two consequences for the map:

  1. The response half of the schema cannot be designed from public sources. It needs a real
    document obtained through a relationship or a records request. That is why this ticket blocks
    the schema ticket.
  2. It is also a point in the idea's favour. Nobody has standardised responses partly because
    nobody can see them at scale. A rail that sits where responses are actually submitted would hold
    a corpus that cannot be assembled any other way — which is a defensible position, not just a
    feature. Worth weighing when the HelmDocs integration boundary is drawn.

Checklist for the human

Ordered by cost. Any one of these unblocks the ticket.

  1. Ask a vendor directly. Fastest by far. Anyone in the network who has responded to a large
    retail or institutional RFP and will share a redacted copy. One real response beats ten
    scraped ones.
  2. Public records request to an issuer with a known award. Tarrant County (#2021-137, jail food
    services) and Friendship PCS both have identifiable awarded vendors and both operate under
    public-records regimes. Request the winning submission, not the tabulation.
  3. State portals with disclosure defaults — Maine's procurement site states awarded submissions
    "will be considered public records." Find the equivalent in a state that posts them rather than
    merely permitting a request.
  4. Buy one. Responsive, Loopio, and Proposify all publish sample/template responses to
    prospects. A template is weaker evidence than a real submission but is better than nothing for
    the binding model specifically.

Done when one complete response is in hand — ideally with a filled compliance matrix — and the
requirement-to-response binding has been checked against a real document. Record what was obtained,
from where, and what it changes.

## Progress — AFK attempt failed, handing to a human Retried the 403'd sources through a real browser rather than plain HTTP. The bot-challenge theory was wrong: **the documents are not blocked, they are not published.** **Searched, nothing found:** - Sourcewell awarded-supplier proposals (`assets.sourcewell-mn.gov` returns zero indexed proposals) - OMNIA Partners / Region 4 supplier responses - Generic hunt for a filled compliance matrix (`"compliance matrix" "RFP section" "comply"`) — every hit was the RFP *instructing* bidders to produce one, never a completed one - Louisiana Medicaid MCO RFP 2021, a $9bn award with six named proposers (Aetna Better Health, AmeriHealth Caritas, et al.) — LDH publishes the RFP and the proposer list, not the proposals **The finding worth keeping.** Solicitations are published because issuers are legally obliged to advertise them. Responses are commercially sensitive and are published essentially never, even by issuers whose own RFP says awarded submissions become public records. There is no public response corpus, in any vertical. Two consequences for the map: 1. **The response half of the schema cannot be designed from public sources.** It needs a real document obtained through a relationship or a records request. That is why this ticket blocks the schema ticket. 2. **It is also a point in the idea's favour.** Nobody has standardised responses partly because nobody can see them at scale. A rail that sits where responses are actually submitted would hold a corpus that cannot be assembled any other way — which is a defensible position, not just a feature. Worth weighing when the HelmDocs integration boundary is drawn. ## Checklist for the human Ordered by cost. Any one of these unblocks the ticket. 1. **Ask a vendor directly.** Fastest by far. Anyone in the network who has responded to a large retail or institutional RFP and will share a redacted copy. One real response beats ten scraped ones. 2. **Public records request** to an issuer with a known award. Tarrant County (#2021-137, jail food services) and Friendship PCS both have identifiable awarded vendors and both operate under public-records regimes. Request the winning submission, not the tabulation. 3. **State portals with disclosure defaults** — Maine's procurement site states awarded submissions "will be considered public records." Find the equivalent in a state that posts them rather than merely permitting a request. 4. **Buy one.** Responsive, Loopio, and Proposify all publish sample/template responses to prospects. A template is weaker evidence than a real submission but is better than nothing for the binding model specifically. **Done when** one complete response is in hand — ideally with a filled compliance matrix — and the requirement-to-response binding has been checked against a real document. Record what was obtained, from where, and what it changes.
Author
Owner

Resolution

Real completed vendor responses could not be obtained, and this ticket is closed unresolved on its
original terms.
Retrieval was exhausted: completed responses are not published anywhere, and
browser retrieval confirmed the earlier 403 theory was wrong — they are not access-blocked, they
simply do not exist in public form. No amount of further searching produces one.

The decision

v1 is reframed as a demoable proof of concept. The common schema will be designed against a
synthesized response corpus, and the map proceeds rather than stalling.

What is actually missing, precisely

This matters for how much the synthesis can be trusted. Collect real retail RFPs and vendor
responses
did land real solicitations, and with them:

  • real scoring scales, which it established are per-document and must never be normalised
  • real price structures, which it established are rate functions, not scalars
  • the response binding gradient — from tabbed restatement of requirements down to free prose

So the RFP half of the schema rests on real artifacts. The gap is one-sided: the response half.
Synthesis therefore means writing responses to solicitations we actually hold, constrained by
their real requirement lists, real scoring rubrics, and real submission instructions — not inventing
both sides of the exchange.

Standard for the synthesis

Lightweight but with teeth. The corpus exists to stress the schema, not to flatter it, so it must
include the cases that break naive models:

  • a response that restates requirements in a tabbed compliance matrix (the easy end of the
    binding gradient) and one that answers in free prose with no structural anchor (the hard end)
  • pricing that is a rate function — tiered, volume-banded, or unit-times-quantity — not a single
    number
  • at least one response that is partially non-responsive, omitting a mandatory requirement,
    since ranking has to represent absence rather than assume completeness
  • exceptions and qualifications taken against the retailer's terms, which are the thing real
    evaluators care about most and the thing an invented corpus always forgets

Accepted risk, recorded

The schema will be under-constrained exactly where a real response would have surprised us
most likely in free-prose binding and in price-table shapes, since those are the two places the
research already flagged as high-variance. A synthesized corpus reflects the schema's author's
assumptions back at them, so it cannot falsify the design, only exercise it.

This is a known and accepted cost of the POC framing. Obtaining even one real completed response
remains the single highest-value validation available
, and it should be revisited before anything
built on this schema is treated as production-ready.

Consequence

Define the common schema for RFPs and responses is unblocked. The synthesized corpus is a
deliverable of that ticket, produced alongside the schema rather than before it, so that each
stresses the other.

## Resolution **Real completed vendor responses could not be obtained, and this ticket is closed unresolved on its original terms.** Retrieval was exhausted: completed responses are not published anywhere, and browser retrieval confirmed the earlier 403 theory was wrong — they are not access-blocked, they simply do not exist in public form. No amount of further searching produces one. ### The decision **v1 is reframed as a demoable proof of concept.** The common schema will be designed against a **synthesized response corpus**, and the map proceeds rather than stalling. ### What is actually missing, precisely This matters for how much the synthesis can be trusted. [Collect real retail RFPs and vendor responses](https://gitea.stephenmann.io/christian/helmdocs-proposal-system/issues/22) *did* land real solicitations, and with them: - real scoring scales, which it established are **per-document and must never be normalised** - real price structures, which it established are **rate functions, not scalars** - the **response binding gradient** — from tabbed restatement of requirements down to free prose So the RFP half of the schema rests on real artifacts. **The gap is one-sided: the response half.** Synthesis therefore means writing responses *to solicitations we actually hold*, constrained by their real requirement lists, real scoring rubrics, and real submission instructions — not inventing both sides of the exchange. ### Standard for the synthesis Lightweight but with teeth. The corpus exists to **stress the schema**, not to flatter it, so it must include the cases that break naive models: - a response that **restates requirements in a tabbed compliance matrix** (the easy end of the binding gradient) and one that **answers in free prose with no structural anchor** (the hard end) - **pricing that is a rate function** — tiered, volume-banded, or unit-times-quantity — not a single number - at least one response that is **partially non-responsive**, omitting a mandatory requirement, since ranking has to represent absence rather than assume completeness - **exceptions and qualifications** taken against the retailer's terms, which are the thing real evaluators care about most and the thing an invented corpus always forgets ### Accepted risk, recorded The schema will be **under-constrained exactly where a real response would have surprised us** — most likely in free-prose binding and in price-table shapes, since those are the two places the research already flagged as high-variance. A synthesized corpus reflects the schema's author's assumptions back at them, so it **cannot falsify the design**, only exercise it. This is a known and accepted cost of the POC framing. **Obtaining even one real completed response remains the single highest-value validation available**, and it should be revisited before anything built on this schema is treated as production-ready. ### Consequence [Define the common schema for RFPs and responses](https://gitea.stephenmann.io/christian/helmdocs-proposal-system/issues/10) is unblocked. The synthesized corpus is a deliverable of that ticket, produced alongside the schema rather than before it, so that each stresses the other.
Sign in to join this conversation.
No description provided.